The PR Checklist We Run on Every AI-Generated Merge
We use AI to write 60% of our code. Here is the five-point checklist we run on every PR because of what happened when we did not.

Author
Co-Founder & Backend Engineer
location_onBudapest, HU
Tamas leads the Python backend, data layer, and security review at CZ Dev. Day job: vulnerability threat management analyst at Citi, where he built the GEM Dashboard in two months, shipped the Firewall and Load Balancer Vulnerability Dashboard solo, and wrote the email composer that cleared Apache Tika XXE and Oracle HTTP Server RCE L0 incidents inside their 48-hour windows, replacing a 4-hour manual mail-merge with a 12-minute pipeline. Before Citi: on-call for critical OpenStack infrastructure at Ericsson (95% on-time delivery, 90% queue utilization on concurrent cases) and reporting under intra-day SLAs at BlackRock (200+ daily reports, 30% direct ticket reduction). At CZ Dev he ships VitalRegistry v1 and v2, KEV Explorer, and Advisory Composer, and leads security review across the portfolio. Stack: Python, TypeScript/React, Firebase. BSc in Business IT, Budapest Business School.
15 Posts
We use AI to write 60% of our code. Here is the five-point checklist we run on every PR because of what happened when we did not.
We built VitalRegistry across Budapest and Finland. Dropped standups, missed reviews, features dark for two weeks. Here is the 3-rule system that fixed it.
CZ Dev did not brainstorm VitalRegistry. We watched a BEMER distributor manage 200 rental devices in Excel for a year, then built what she actually needed.
When a real rental operation outgrows spreadsheets and every no-code workaround costs more than a custom build would have, this is what we built instead.
I built two products I'm proud of with a $20 Claude subscription. The hard part was never the code.
We built czaban.dev with Claude. It looked right. The /seo audit scored it 59/100. Here is every issue it found and how we fixed them in one day.
Three operational patterns that signal your internal tool has crossed the break-even point on a rebuild, before the costs become obvious.
A decision framework for early-stage founders built from real scaling choices, not theory. Know when Streamlit breaks before it does.
The six-month-silence failure mode isn't a talent problem. It's a discovery process that never defined done. Here's what to ask in week one.
A transparent pricing breakdown using a real CRM build: what drove cost up, what saved it, and why the cost of staying put is the number founders forget to calculate.
A full case study of VitalRegistry: the decisions, the rewrite trigger, and what ten real users taught us about custom tooling.
The financial break-even for switching from no-code to custom software arrives sooner than founders expect. The real cost is coordination overhead, not subscription price.
We audited five small software agency websites to understand what they were doing with SEO. The findings were consistent. Most are invisible to the searches that should find them.
No-code platforms are a legitimate starting point. But there's a predictable moment when they stop working, and most founders wait too long to notice it.
How we turned a 47-column Google Sheet and three email threads into a Streamlit app that founders actually want to use.